DAIOE: Dynamic AI Occupational Exposure
Abstract
DAIOE translates measured performance gains on public AI benchmarks into occupation-year exposure scores, separately for nine capability subdomains and a generative-AI composite. Unlike static exposure indices it varies over time as well as across occupations, so the timing of capability arrival is observable and testable. This record contains the occupation-year scores on five occupational classifications (O*NET-SOC 2010, SOC 2010, ISCO-08, SSYK 96, SSYK 2012) in Stata, TSV and Excel formats, for two vintages held separately: the frozen 2010-2023 index behind the published estimates, and the 2024 refresh. A SOC 2018 build on the frozen window ships alongside. The vintages are not interchangeable, and the frozen index is the one to use for replication; cite the vintage you used. The 2025-onward vintage will follow as a later version of this record. The measure is built from public AI benchmark results and O*NET occupational ability profiles. Two properties are documented rather than left to be discovered: benchmarks enter and retire as research moves, so an application's basket thins once its benchmarks are solved, and the subdomain series are strongly correlated with one another, which limits how far the decomposition can attribute an effect to any single capability. The archive includes the licence terms for the data, a provenance file listing every measurement recorded from Papers with Code together with the paper that first published it, and SHA-256 checksums for all files. The construction pipeline, its test suite and full technical documentation are at https://github.com/Magnus-L/daioe-pipeline Code is MIT licensed; the scores are CC BY 4.0.
